\begin{table}[t]
\caption{Synthesis of the five research questions.}\label{tab:rq_synthesis}
\small
\setlength{\tabcolsep}{3pt}
\begin{tabular}{@{}L{0.07\textwidth}L{0.35\textwidth}L{0.27\textwidth}L{0.21\textwidth}@{}}
\toprule
RQ & Main answer & Primary evidence & Boundary or caveat \\
\midrule
RQ1 & Modern noisy supervision includes semantic ambiguity, ontology mismatch, open-set contamination, weak targets, and process-dependent labels that cannot be reduced to random flips. & Taxonomy, structured supervision mismatch, and recoverability analysis & Different mismatches imply different latent targets; a single hidden clean label may not exist. \\
RQ2 & Reliability becomes observable when optimization exposes stable differences in difficulty, trajectories, or representation geometry; it disappears when hard clean and coherent noisy samples overlap. & Memorization dynamics, representation diagnostics, and the small-loss case study & Observability depends on class balance, semantic structure, model state, and information budget. \\
RQ3 & Robust methods can be compared through five interacting roles: characterization, isolation, refinement, representation preservation, and optimization or uncertainty control. & Unified pipeline and assumption-aware method guide & A component's role and effectiveness can change with the noise process and available priors. \\
RQ4 & A benchmark supports a robustness claim only when its noise-generating process, available clean information, metrics, and stress tests match the claimed setting. & Protocol comparison, benchmark case studies, and reporting checklist & Final accuracy alone cannot reveal hard-clean rejection, calibration failure, or representation damage. \\
RQ5 & Emerging settings change the unit of noise, the source of reliability, and the evaluation target, thereby stress-testing assumptions inherited from closed-set classification. & Vision, foundation-model, federated, graph, and domain-specific evidence & Recent and preprint evidence should be treated as directional unless supported across settings. \\
\botrule
\end{tabular}
\end{table}